On the two-times differentiability of the value functions ...

The Annals of Applied Probability
2006, Vol. 16, No. 3, 1352–1384
DOI: 10.1214/105051606000000259
© Institute of Mathematical Statistics, 2006

ON THE TWO-TIMES DIFFERENTIABILITY OF THE VALUE
FUNCTIONS IN THE PROBLEM OF OPTIMAL
INVESTMENT IN INCOMPLETE MARKETS

BY DMITRY KRAMKOV1 AND MIHAI SÎRBU

Carnegie Mellon University and Columbia University

We study the two-times differentiability of the value functions of the pri-
mal and dual optimization problems that appear in the setting of expected util-
ity maximization in incomplete markets. We also study the differentiability of
the solutions to these problems with respect to their initial values. We show
that the key conditions for the results to hold true are that the relative risk
aversion coefficient of the utility function is uniformly bounded away from
zero and infinity, and that the prices of traded securities are sigma-bounded
under the numéraire given by the optimal wealth process.

1. Introduction and main results. We study a similar financial framework
to the one in [7] and refer to this paper for more details and references. We con-
sider a model of a security market which consists of d + 1 assets, one bond and
d stocks. We work in discounted terms, that is, we suppose that the price of the
bond is constant and denote by S = (Si)1≤i≤d the price process of the d stocks.
The process S is assumed to be a semimartingale on a filtered probability space
( , F , (Ft )0≤t≤T , P). Here T is a finite time horizon and F = FT .

A (self-financing) portfolio is defined as a pair (x, H ), where the constant x rep-
resents the initial capital and H = (H i)1≤i≤d is a predictable S-integrable process,
Hti specifying how many units of asset i are held in the portfolio at time t. The
wealth process X = (Xt )0≤t≤T of the portfolio evolves in time as the stochastic
integral of H with respect to S:

t

(1) Xt = X0 + Hu dSu, 0 ≤ t ≤ T .

0

We denote by X(x) the family of nonnegative wealth processes with initial
value x:

(2) X(x) = {X ≥ 0 : X is defined by (1) with X0 = x}.

A probability measure Q ∼ P is called an equivalent local martingale measure
if any X ∈ X(1) is a local martingale under Q. The family of equivalent local

Received February 2005; revised December 2005.
1Supported in part by the NSF Grants DMS-01-39911 and DMS-05-05414.
AMS 2000 subject classifications. Primary 90A09, 90A10; secondary 90C26.
Key words and phrases. Utility maximization, incomplete markets, Legendre transformation, du-
ality theory, risk aversion, risk tolerance.

1352

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1353

martingale measures is denoted by Q. We assume throughout that

(3) Q = ∅.

This condition is intimately related to the absence of arbitrage opportunities on the
security market. See [1] for precise statements and references.

We also consider an economic agent in our model, whose preferences over ter-
minal wealth are modeled by a utility function U = (U (x))x>0. The function U is
assumed to be strictly concave, strictly increasing and continuously differentiable
and to satisfy the Inada conditions:

(4) U (0) = lim U (x) = ∞, U (∞) = lim U (x) = 0.
x→0 x→∞

In what follows we set U (0) = limx→0 U (x) and U (x) = −∞ for all x < 0.
For a given initial capital x > 0, the goal of the agent is to maximize the expected

value of terminal utility. The value function of this problem is denoted by

(5) u(x) = sup E[U (XT )].

X∈X(x)

Intuitively speaking, the value function u = (u(x))x>0 plays the role of the utility
function of the investor at time 0, if he/she subsequently invests in an optimal way.
To exclude the trivial case, we assume that u is finite:

(6) u(x) < ∞, x > 0.

A well-known tool in studying the optimization problem (5) is the use of duality
relationships in the spaces of convex functions and semimartingales. Following [7],
we define the dual optimization problem to (5) as follows:

(7) v(y) = inf E[V (YT )], y > 0.

Y ∈Y(y)

Here V is the convex conjugate function to U , that is,

V (y) = sup{U (x) − xy}, y > 0,

x>0

and Y(y) is the family of nonnegative supermartingales Y that are dual to X(1) in
the following sense:

(8) Y(y) = {Y ≥ 0 : Y0 = y and XY is a supermartingale for all X ∈ X(1)}.

Note that the set Y(1) contains the density processes of all Q ∈ Q.
The optimization problems (5) and (7) are well studied. For example, it was

shown in [7] that the value functions u and v are conjugate, that is,

(9) v(y) = sup{u(x) − xy}, y > 0,

x>0

and that the minimal market independent condition on U that implies the contin-
uous differentiability of the value functions u and v on (0, ∞) and the existence

1354 D. KRAMKOV AND M. SÎRBU

of the solutions X(x) and Y (y) to (5) and (7) for all x > 0 and y > 0 is that the
asymptotic elasticity of U is strictly less than 1, that is,

(10) lim sup xU (x) < 1.

x→∞ U (x)

In addition, if y = u (x) then

(11) U (XT (x)) = YT (y)

and the product X(x)Y (y) is a martingale. Hereafter, we shall use these results
without further comment.

In this paper we are interested in the existence and the computation of the
second derivatives u (x) and v (y) of the value functions u and v and of the
first derivatives X (x) and Y (y) of the solutions X(x) and Y (y) with respect to
x and y. In addition to a purely theoretical interest (see Remark 1 below) our study
of these questions is also motivated by some applications. For example, in [9] we
use the results of the current paper to perform the sensitivity analysis of utility-
based prices with respect to the number of nontraded contingent claims.

REMARK 1. If S is a Markov diffusion process, then (5) becomes a typical
stochastic control problem and can be studied using PDE methods. In this case, the
two-times differentiability of the value function u is closely related to the existence
of the classical solution of the corresponding Bellman equation. We refer to [10]
for the deep treatment of this topic.

To give positive answers to the above questions we need to impose additional
conditions on the utility function U and the price process S. The conditions on U
are stated in the following assumption:

ASSUMPTION 1. The utility function U is two-times differentiable on (0, ∞)
and its relative risk aversion coefficient

(12) A(x) = − xU (x) x > 0,
,
U (x)

is uniformly bounded away from zero and infinity, that is, there are constants
c1 > 0 and c2 < ∞ such that

(13) c1 < A(x) < c2, x > 0.

In Lemma 3 below we prove that the bounds (13) on the relative risk aversion
coefficient imply both the Inada conditions (4) and the condition (10) on the as-
ymptotic elasticity. Note that U is two-times differentiable at x > 0 and U (x) < 0
if and only if the conjugate function V is two-times differentiable at y = U (x),
and that in this case

(14) V (y) = − 1 .
U (x)

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1355

It follows that (13) is equivalent to the following condition:

(15) 11 y > 0,
where < B(y) < ,

c2 c1

(16) B(y) = − yV (y) , y > 0.

V (y)

Note that if y = U (x) then

(17) B(y) = 1
A(x)

is the relative risk tolerance coefficient of U computed at x.
To facilitate the formulation of the assumption on the price process S we give

the following definition:

DEFINITION 1. A d-dimensional semimartingale R is called sigma-bounded
if there is a strictly positive predictable (one-dimensional) process h such that the
stochastic integral h dR is well defined and is locally bounded.

This definition has been motivated by a similar concept of sigma-martingales
which plays the key role in the fundamental theorem of asset pricing for unbounded
processes; see [1] for the details. Both these notions are instances of a more general
concept of sigma-localization studied in [6]. Appendix A below contains a number
of equivalent reformulations of Definition 1, as well as other useful results on
sigma-bounded semimartingales and martingales.

In principle, any strictly positive wealth process X can be chosen as a new
currency or a numéraire in the model. In this case, the (d + 1)-dimensional semi-
martingale

(18) SX = 1S
,
XX

has the economic interpretation as the prices of traded securities (the bond and the
stocks) discounted by X. It is well known that a process X is a stochastic integral
with respect to SX if and only if XX is a stochastic integral with respect to S.
In other words, X is a wealth process under the numéraire X if and only if XX
is a wealth process in the original financial model, where the role of the money
denominator is played by the bond. We shall use this fact on several occasions.

Hereafter, we fix an initial capital x > 0 and denote y = u (x). The following
assumption plays a key role in the proofs of our main results. It states that the
prices of traded securities are sigma-bounded under the numéraire given as the
solution X(x) to (5).

1356 D. KRAMKOV AND M. SÎRBU

ASSUMPTION 2. The price process of the traded securities discounted
by X(x), that is, the (d + 1)-dimensional semimartingale

(19) SX(x) = 1S
,
X(x) X(x)

is sigma-bounded.

The direct verification of Assumption 2 is feasible only if we can compute the
optimal wealth process X(x) explicitly. In other cases, some sufficient “qualita-
tive” conditions on the financial model should be used. An example of such a
condition is given by Theorem 3 in Appendix A. This theorem shows that all
semimartingales defined on the filtered probability space ( , F , (Ft )0≤t≤T , P)
are sigma-bounded (and, hence, Assumption 2 holds true trivially) if the family of
purely discontinuous martingales admits a finite-dimensional basis (from the point
of view of stochastic integration). In particular, as Proposition 2 in Appendix A
shows, all semimartingales are sigma-bounded if the financial model is complete
(or can be extended to become complete by adding to it a finite number of addi-
tional securities). It is interesting to note that there are complete financial models
where the “locally bounded” version of Assumption 2 fails to hold true; see Ex-
ample 5 in Section 4.

To state our main results we also need to define two auxiliary optimization prob-
lems. Let R(x) be the probability measure whose Radon–Nikodym derivative un-
der P is given by

(20) dR(x) = XT (x)YT (y) = XT (x)U (XT (x)) .
dP xy xu (x)

(A similar change of measure was used in the particular case U (x) = −x2 in [2].)
Let H20(R(x)) be the space of square integrable martingales under R(x) with ini-
tial value 0. We would like to point out that square integrability is naturally related

to the existence of second-order derivatives, as the reader can see below. Denote
by M2(x) the subspace of H02(R(x)) that consists of stochastic integrals with re-
spect to SX(x), that is,

M2(x) = M ∈ H02(R(x)) : M = H dSX(x) for some H .

Further, let N 2(y) = N 2(u (x)) be the orthogonal complement of M2(x) in
H20(R(x)). In other words, N ∈ N 2(y) if and only if N ∈ H20(R(x)) and MN
is a martingale under R(x) for all M ∈ M2(x).

After these preparations, we formulate the following optimization problems:

(21) a(x) = inf ER(x)[A(XT (x))(1 + MT )2],

M ∈M2 (x )

(22) b(y) = inf ER(x)[B(YT (y))(1 + NT )2], y = u (x),

N ∈N 2(y)

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1357

where the functions A and B are defined in (12) and (16), respectively. The basic
properties of these optimization problems are stated in Lemma 1 below. The proof
of this lemma will be given in Section 3 and will follow from its “abstract version,”
Lemma 2.

LEMMA 1. Assume that conditions (3), (6) and Assumption 1 hold true. Let
x > 0 and y = u (x). Then:

1. The value functions a(x) and b(y) defined in (21) and (22) satisfy

a(x)b(y) = 1

and

c1 < a(x) < c2, 11
< b(y) < ,

c2 c1

where the constants c1 and c2 appear in (13).
2. The solutions M(x) to (21) and N(y) to (22) exist and are unique. In addition,

(23) A(XT (x)) 1 + MT (x) = a(x) 1 + NT (y) .

For x > 0 and y = u (x) we define the semimartingales X (x) and Y (y):

(24) X (x) = X(x) 1 + M(x) ,
x

(25) Y (y) = Y (y) 1 + N(y) ,
y

where M(x) and N(y) are the solutions to (21) and (22), respectively. Note that as
M(x) is a stochastic integral with respect to SX(x), the semimartingale X (x) is a
stochastic integral with respect to S. In other words, X (x) is a wealth process.

The following theorem is the main result of our paper:

THEOREM 1. Let x > 0 and denote y = u (x). Assume that conditions
(3) and (6) and Assumptions 1 and 2 hold true. Then:

1. The second derivatives u (x) and v (y) of the value functions u and v defined
in (5) and (7) exist at x and y, respectively, and

xu (x)
(26) c1 < − u (x) = a(x) < c2,

(27) 1 < −yv (y) = b(y) < 1
,
c2 v (y) c1

where a(x) and b(y) are the value functions defined in (21) and (22), and the
constants c1 and c2 are given in (13).

1358 D. KRAMKOV AND M. SÎRBU

2. The derivatives of the terminal values of the solutions X(x) to (5) and Y (y)
to (7) with respect to x and y exist and equal the terminal values of the semi-
martingales X (x) and Y (y) defined in (24) and (25). That is,

(28) lim XT (x + ε) − XT (x) = XT (x),
ε→0 ε

(29) lim YT (y + ε) − YT (y) = YT (y),
ε→0 ε

where the convergence takes place in P-probability.
3. We have

(30) u (x) = E[U (XT (x))(XT (x))2],
(31) v (y) = E[V (YT (y))(YT (y))2].

Moreover,

(32) U (XT (x))XT (x) = u (x)YT (y),

and the products X(x)Y (y), X (x)Y (y) and X (x)Y (y) are martingales un-
der P.

REMARK 2. According to (28), the wealth process X (x) describes how the
agent invests a small additional unit of capital. Note that, while the construction
of the processes X (x) and Y (y) by (24) and (25) was based on Lemma 1 and,
hence, did not require Assumption 2, the equalities (28) and (29) hold true, in gen-
eral, only if this sigma-boundedness assumption is satisfied. Note also that (under
the conditions of Theorem 1) equalities (28) and (29) and the martingale proper-
ties of X (x)Y (y) and X(x)Y (y), where y = u (x), allow us to compute X (x)
and Y (y) directly from X(x) and Y (y) without relying on the optimization prob-
lems (21) and (22). Finally, we point out that, contrary to a naive conjecture, the
processes X (x) and Y (y) might not be positive; see Example 4 in Section 4 be-
low.

The proof of Theorem 1 will be given in Section 3 after we study the “abstract”
version of this theorem in Section 2. In Section 4 we construct counterexamples to
some natural but false conjectures related to our main results. In particular, these
counterexamples show that, in general, the upper and the lower bounds in (13), as
well as Assumption 2, cannot be removed without affecting the existence of the
second derivatives u and v . Finally, in Appendix A we present some results on
sigma-bounded semimartingales and martingales. In particular, Theorem 3 there
contains convenient sufficient conditions on the underlying filtered probability
space ( , F , (Ft )0≤t≤T , P) that ensure each semimartingale X is sigma-bounded
and, hence, guarantee the validity of Assumption 2.

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1359

2. The abstract versions of the main results. Hereafter, we use the standard
notation L0 for the set of all random variables and L∞ for the set of bounded
random variables on ( , F , P). If Q ∼ P, then we denote
(33) L02(Q) = {g ∈ L0 : EQ[g] = 0 and EQ[g2] < ∞}.

We start with the abstract version of Lemma 1. Let A and B be nonempty
complementary linear subspaces of L20(P), that is,

α ∈ A ⇐⇒ α ∈ L02(P) and E[αβ] = 0 ∀ β ∈ B,
(34)

β ∈ B ⇐⇒ β ∈ L02(P) and E[αβ] = 0 ∀ α ∈ A.

Let ζ be a random variable such that

(35) c1 < ζ < c2,

for some constants 0 < c1 < c2 < ∞, and let η be the reciprocal to ζ :
1

η= .
ζ

We consider the following optimization problems:
(36) a = inf E[ζ (1 + α)2],

α∈A

(37) b = inf E[η(1 + β)2].

β∈B

LEMMA 2. Assume (34) and (35). Then:

1. The numbers a and b defined in (36) and (37) satisfy

(38) ab = 1

and

(39) 11
c1 < a < c2, <b< ,
c2 c1

where the constants c1 and c2 appear in (35).
2. The solutions α to (36) and β to (37) exist and are unique. In addition,

(40) ζ (1 + α) = a(1 + β).

PROOF. First, inequalities (39) are easy consequences of (35). Further, let
(γn)n≥1 be a sequence in A such that

lim E[ζ (1 + γn )2 ] = a.

n→∞

As ζ ≥ c1, the sequence (γn)n≥1 is bounded in L02(P). Hence, there is a sequence
of convex combinations

αn ∈ conv(γn, γn+1, . . .), n ≥ 1,

1360 D. KRAMKOV AND M. SÎRBU

that converges in L02(P) to some α. As A is closed in L02(P), we have α ∈ A. The
convexity of (1 + x)2 and the inequality ζ ≤ c2 now imply that

E[ζ (1 + α)2] = lim E[ζ (1 + αn)2] ≤ lim sup E[ζ (1 + γn)2] = a .

n→∞ n→∞

This proves that α is a solution to (36). The fact that α is the only solution to (36)
follows from the strict convexity of (1 + x)2.

Using standard arguments from the calculus of variations, we deduce that the

optimality of α implies that for any α ∈ A,

E[ζ (1 + α)α] = 0.

From the complementary relations (34) between A and B, we deduce the existence
of a constant c and a random variable γ ∈ B such that

ζ (1 + α) = c + γ .

If we multiply both sides of this equality by 1 + α and compute the expected value,
we get c = a = 0. Hence, denoting β = γ /a, we deduce that

(41) ζ (1 + α) = a(1 + β),

where β ∈ B.
Repeating the same arguments for the optimization problem (37), we deduce

the existence and the uniqueness of the solution β to this problem, as well as the
representation

(42) η(1 + β) = b(1 + α),

for some α ∈ A.
By multiplying (41) and (42), and using the fact that ζ η = 1, we arrive at the

equality

(1 + α)(1 + β) = ab(1 + α)(1 + β),

which, after taking the expectation under P, implies the relation (38) between
a and b. The equality of β defined in (41) to β now follows from the uniqueness
of the solution to (37) and the computations below:

E[η(1 + β)2] = E η ζ2 (1 + α)2 = 1 E[ζ (1 + α)2] = 1 = b.
a2 a2 a

In the remaining part of this section we state and prove the abstract version of
Theorem 1. Let C and D be nonempty sets of nonnegative random variables such
that:

1. The set C is bounded in L0 and contains the constant function g = 1:

(43) lim sup P[|g| ≥ n] = 0,

n→∞ g∈C

(44) 1 ∈ C.

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1361

2. The sets C and D satisfy the bipolar relations ∀h ∈ D,
∀g ∈ C.
g ∈ C ⇐⇒ E[gh] ≤ 1
(45)

h ∈ D ⇐⇒ E[gh] ≤ 1

For x > 0 and y > 0, we define the sets

C(x) = xC = {xg : g ∈ C},
D(y) = yD = {yh : h ∈ D},

and the optimization problems

(46) u(x) = sup E[U (g)],

g∈C(x)

(47) v(y) = inf E[V (h)].

h∈D (y )

Here U and V are the functions defined in Section 1.
Hereafter, we assume that

(48) u(x) < ∞, x > 0,

and that Assumption 1 holds true. From Lemma 3 below and Theorem 3.2 in [7]
we deduce that the value functions u and v defined in (46) and (47) are conjugate,
that is, (9) holds true, u and v are continuously differentiable on (0, +∞) and the
solutions g(x) to (46) and h(y) to (47) exist for all x > 0 and y > 0. In addition, if
y = u (x), then

(49) U (g(x)) = h(y),
(50) E[g(x)h(y)] = xy.

Hereafter, we shall use these results without further comment. As in the previous
section, we are interested in the existence of second derivatives u (x) and v (y)
of the value functions and the first derivatives g (x) and h (y) of the solutions to
these problems.

For x > 0 we denote by A∞(x) the family of bounded random variables α such
that g(x)(1 + cα) and g(x)(1 − cα) belong to C(x) for some c = c(α) > 0, where
g(x) is the solution to (46). In other words,

(51) A∞(x) = {α ∈ L∞ : g(x)(1 ± cα) ∈ C(x) for some c > 0}.

Similarly, for y > 0 we denote
(52) B∞(y) = {β ∈ L∞ : h(y)(1 ± cβ) ∈ D(y) for some c > 0},

where h(y) is the solution to (47).

1362 D. KRAMKOV AND M. SÎRBU

Hereafter, we fix x > 0 and denote y = u (x). Let R(x) be the probability mea-
sure on ( , F ) whose Radon–Nikodym derivative under P is given by

(53) d R(x ) = g(x)h(y)
dP .

xy

From the bipolar relations (45) for the sets C and D we deduce that the sets A∞(x)
and B∞(y) = B∞(u (x)) are orthogonal linear subspaces in L20(R(x)). That is,
(54) ER(x)[α] = ER(x)[β] = ER(x)[αβ] = 0 ∀ α ∈ A∞(x), β ∈ B∞(y).

Denote by A2(x) and B2(y) the respective closures of A∞(x) and B∞(y)
in L20(R(x)). From (54) we deduce that A2(x) and B2(y) are closed orthogonal
linear subspaces in L02(R(x)). It turns out that the two-times differentiability of u
at x and of v at y depends crucially on the condition that these two subspaces are

complementary to each other.

ASSUMPTION 3. The sets A2(x) and B2(y), where y = u (x), are comple-
mentary linear subspaces in L02(R(x)). That is,

α ∈ A2(x) ⇐⇒ α ∈ L02(R(x)) and ER(x)[αβ] = 0

∀ β ∈ B2(y),
(55)

β ∈ B2(y) ⇐⇒ β ∈ L20(R(x)) and ER(x)[αβ] = 0

∀ α ∈ A2(x).

As we show in Section 3, Assumption 3 is the “abstract version” of Assump-
tion 2.

Consider now the optimization problems
(56) a(x) = inf ER(x)[A(g(x))(1 + α)2],

α∈A2(x)

(57) b(y) = inf ER(x)[B(h(y))(1 + β)2], y = u (x),

β ∈B 2 (y )

where the functions A and B are defined in (12) and (16). From Lemma 2 we
deduce that if Assumptions 1 and 3 hold true, then the solutions α(x) to (56)
and β(y) to (57) exist and are unique, and [recalling that y = u (x)]

(58) a(x)b(y) = 1,

(59) A(g(x)) 1 + α(x) = a(x) 1 + β(y) .

Using this notation, we define the random variables
(60) g (x) = g(x) 1 + α(x) ,

x
(61) h (y) = h(y) 1 + β(y) .

y

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1363

THEOREM 2. Let x > 0 and denote y = u (x). Assume that conditions
(43)–(45), (48) and Assumptions 1 and 3 hold true. Then:

1. The value functions u and v defined in (46) and (47) are two-times differentiable
at x and y, respectively, and

(62) c1 < − xu (x) = a(x) < c2,
u (x)

1 yv (y) 1
(63) < − = b(y) < ,
c2 v (y) c1

where a(x) and b(y) are defined in (56) and (57) and the constants c1 and c2
are given in (13).
2. The derivatives of the solutions g(x) to (46) and h(y) to (47) with respect to x
and y exist and equal g (x) and h (y), respectively, as defined in (60) and (61),
that is,

g(x + ε) − g(x)
(64) lim = g (x),

ε→0 ε
h(y + ε) − h(y)

(65) lim = h (y),
ε→0 ε

where the convergence takes place in P-probability.
3. We have

(66) u (x) = E[U (g(x))(g (x))2],
(67) v (y) = E[V (h(y))(h (y))2].

Moreover,

(68) U (g(x))g (x) = u (x)h (y).

Before proceeding to the proof of the theorem, we state two technical results
related to the condition (13) on the utility function U .

LEMMA 3. Assume that the utility function U = U (x) satisfies (13). Then the
following assertions hold true:

1. For any constant a > 1 there are constants 0 < b1 < b2 < 1 such that

(69) b1U (x) < U (ax) < b2U (x), x > 0.

2. The function U satisfies the Inada conditions (4).
3. The asymptotic elasticity of U is strictly less than 1, that is, (10) holds true.

1364 D. KRAMKOV AND M. SÎRBU

PROOF. Let c1 and c2 be the constants defined in (13). Without any loss of
generality we can assume that a > 1 is sufficiently close to 1 so that

1 − c2 ln a > 0.

Using (13) we deduce

a a U (tx)
U (x) − U (ax) = −xU (tx) dt ≥ c1 t dt > c1U (ax) ln a,

11

U (x) − U (ax) = a (tx) dt ≤ c2 a U (tx)
1 t dt < c2U (x) ln a.
−xU

1

The inequalities in (69) now follow, with

b1 = 1 − c2 ln a, b2 = 1 + 1 ln a .
c1

The Inada conditions (4) follow from (69):

U (∞) = lim U (an) ≤ nl→im∞(b2)nU (1) = 0,

n→∞

U (0) = lim U 1 ≥ lim 1 n
an
n→∞ n→∞ b2 U (1) = ∞.

Finally, the fact that the second inequality in (69) implies (10) has been proved

in [7], Lemma 6.5.

LEMMA 4. Assume that the utility function U = U (x) satisfies (13). Let ζ be
a strictly positive random variable such that

(70) E[|U (ζ )|] < ∞

and η a random variable such that

|η| ≤ Kζ

for some K > 0. Then the function

w(s) = E[U (ζ + sη)]

is well defined and two-times differentiable for |s| < 1 . Furthermore,
K

(71) w (s) = E[U (ζ + sη)η], w (s) = E[U (ζ + sη)η2].

PROOF. First, we show that
(72) E[U (ζ )ζ ] < ∞.

Fix a > 1. From (70) and the fact that U is an increasing concave function we
deduce that

E[|U (aζ )|] < ∞.

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1365

It follows that

E[U (aζ )ζ ] < 1 E[U (aζ ) − U (ζ )] < ∞,
a−1

which, together with (69), implies (72).

Further, let 0 < b < 1 . From (13) and (69), we deduce the existence of c > 0
K
such that, for all |s| ≤ b,

|U (ζ + sη)η| + |U (ζ + sη)η2| ≤ cU (ζ )ζ

and, therefore, for all |s| ≤ b and |t| ≤ b,

U (ζ + sη) − U (ζ + tη) + U (ζ + sη) − U (ζ + tη) ≤ 2cU (ζ )ζ.
s−t s−t η

The assertion of the lemma now follows from (72) and the Lebesgue theorem on
dominated convergence.

PROOF OF THEOREM 2. As in the statement of the theorem, we fix x > 0 and
denote y = u (x). We start with the assertions of item 1. Denote

(73) φ(x) = − u (x) a(x),
x

(74) ψ (y ) = −v (y)
b(y).
y

From (58), we deduce that

(75) φ(x)ψ(y) = −1, y = u (x).

We claim that

(76) u(x + ε) ≥ u(x) + u (x)ε + 1 φ (x)ε2 + o(ε2),
2

(77) v(y + ε) ≤ v(y) + v (y)ε + 1 ψ (y)ε2 + o(ε2),
2

where we used the standard generic notation o(ε) for any function f such that
limε→0 f (ε)/|ε| = 0.

If α ∈ A∞(x) then the function

w(s) = E U g(x) 1 + s (1 + α) ,
x

is well defined for sufficiently small s and, by Lemma 4,

(78) w (0) = 1 E[U (g(x))g(x)(1 + α)] = u (x)ER(x)[1 + α] = u (x),
x

w (0) = 1 E[U (g(x))g2(x)(1 + α)2]
x2
(79)
u (x)
= − x ER(x)[A(g(x))(1 + α)2 ],

1366 D. KRAMKOV AND M. SÎRBU

where A is the relative risk aversion coefficient of U defined in (12) and R(x) the

probability measure introduced in (53). Since g(x)(1 + s (1 + α)) ∈ C(x + s) we
x
have

u(x + s) ≥ w(s).

It follows that

lim inf u(x + s) − u(x) − u (x)s

s→0 s2

≥ lim inf w(s) − w(0) − w (0)s
s2
s→0

= 1 (0) = − u (x ) ER(x)[A(g(x ))(1 + α)2].
w 2x

2

Taking sup with respect to α ∈ A∞(x) on the right-hand side of this inequality we

deduce that

lim inf u(x + s) − u(x) − u (x)s ≥ −u (x) = 1 φ (x ),
a(x)
s→0 s2 2x 2

thus proving (76). The proof of (77) is very similar and is omitted here.
Given (76) and (77), the assertions of item 1 are implied by the following result

from convex analysis:

LEMMA 5. Assume that concave functions u and −v are continuously dif-
ferentiable on (0, ∞) and satisfy the conjugacy relations (9). Let x > 0 and
y = u (x). Assume that there are constants φ(x) and ψ(y) that satisfy (75), (76)
and (77).

Then, the functions u and v are two-times differentiable at x and y, and their
respective second derivatives equal φ(x) and ψ(y).

PROOF. For sufficiently small ε we deduce from (9) and (77) that

u(x + ε) ≤ v y + εφ(x) + (x + ε) y + εφ(x)

≤ v(y) + v (y )φ (x )ε + 1 ψ (y)(φ (x)ε)2 + (x + ε) y + εφ(x) + o(ε2)
2

= u(x) + u (x)ε + 1 φ (x )ε2 + o(ε2),
2

where at the last step we used (75) and the equality u(x) = v(y) + xy.
The last inequality and (76) imply that the function u has the following quadratic

expansion at x > 0:

u(x + ε) = u(x) + u (x)ε + 1 φ (x)ε2 + o(ε2).
2

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1367

It is well known (see, e.g., Theorem 5.1.2 in [4]) that for a concave function u the
existence of the quadratic expansion at x is equivalent to the two-times differen-
tiability at this point, and that in this case,

u (x) = φ(x).

Finally, from (9) and (75), we deduce that v is two-times differentiable at
y = u (x), and

v (y) = − 1 = − 1 = ψ(y).
u (x) φ(x)

The assertions of item 3 are straightforward. Indeed, given the definitions of the
random variables g (x) in (60) and h (y) in (61), the representations (66) and (67)
of the second derivatives of u and v are just reformulations of the equalities in
(62) and (63). Further, for y = u (x), the relation (68) between g (x) and h (y)
easily follows from the relation (59) between α(x) and β(y).

We proceed to the proof of the assertions of item 2. First, we note that conver-
gences (64) and (65) are equivalent. Indeed, given, for example, (64) we deduce
from the previous results that

lim h(y + ε) − h(y) = lim h(u (x + ε)) − h(u (x))

ε→0 ε ε→0 u (x + ε) − u (x)

= lim U (g(x + ε)) − U (g(x)) ε

ε→0 ε u (x + ε) − u (x)

= U (g(x))g (x) 1 = h (y).
u (x)

Now let (εn)n≥1 be a sequence of real numbers converging to 0. To complete
the proof of the theorem it remains to be shown that

(80) lim g(x + εn) − g(x) = g (x)
n→∞ εn

in probability. By Lemma 3.6 in [7], we have that

(81) lim g(x + εn) = g(x )

n→∞

in probability. In the future it will be convenient for us to assume that, in fact, the

convergence in (81) takes place almost surely. Of course, this additional assump-

tion does not restrict any generality.

Define the random variables

ζ = 1 min g(x), inf g(x + εn) ,
2
n≥1

η = 2 max g(x), sup g(x + εn) ,

n≥1

θ = 1 inf |U (t )|.
2
ζ ≤t≤η

1368 D. KRAMKOV AND M. SÎRBU

Since the convergence in (81) takes place almost surely, and g(x) > 0, we have

0 < ζ < η < ∞.

From these inequalities and condition (13) on U we deduce that θ is a strictly

positive random variable.
Let (αm)m≥1 be a sequence in A∞(x) that converges to the solution α(x) of (56)

in L20(R(x)) and denote

g(x) m ≥ 1.
gm = x (1 + αm),

Using the fact that A(g(x)) ≤ c2, we deduce

lim ER(x) [A(g(x ))(1 + αm)2] = ER(x) A(g(x)) 1 + α(x) 2 = a(x)

m→∞

and, hence, by the definition of R(x),

(82) lim E[U (g(x))gm2 ] = u (x).

m→∞

For m ≥ 1, denote by n0(m) a sufficiently large integer such that, for all
n ≥ n0(m),

(83) g(x) + εngm ∈ C(x + εn),

(84) |εn|(1 + |αm|) ≤ x .
2

For n ≥ n0(m), we deduce from the Taylor formula and the definition of the ran-
dom variable θ that

U g(x) + εngm − U g(x + εn) ≤ U g(x + εn) g(x) + εngm − g(x + εn)
− θ g(x) + εngm − g(x + εn) 2.

The duality relations (49)–(50) and condition (83) imply that

E U g(x + εn) g(x) + εngm − g(x + εn) ≤ 0.

It follows that

g(x + εn) − g(x) 2 u(x + εn) − E[U (g(x) + εngm)] .
εn εn2
Eθ − gm ≤

By Lemma 4 we have

E U g(x) + εgm = u(x) + u (x)ε + 1 E[U (g(x))gm2 ]ε2 + o(ε2).
2

Combining this quadratic expansion with convergence (82) we get

ml→im∞ nl→im∞ u(x + εn) − E[U (g(x) + εngm)] = 0.
εn2

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1369

Hence,

g(x + εn) − g(x) 2
εn
lim lim sup E θ − gm = 0,

m→∞ n→∞

and the result (80) follows from the facts that θ is a strictly positive random vari-

able and that the sequence (gm)m≥1 converges to g (x) in probability.
The proof of Theorem 2 is complete.

3. Proofs of the main results. We start with the proof of Lemma 1.

PROOF OF LEMMA 1. The proof is an easy consequence of Lemma 2.
Fix x > 0, denote y = u (x) and let A2(x) and B2(y) be the sets of the final
values of the elements from M2(x) and N 2(y), respectively. That is,
(85) A2(x) = {α ∈ L0 : α = MT for some M ∈ M2(x)},
(86) B2(y) = {β ∈ L0 : β = NT for some N ∈ N 2(y)}.
Since M2(x) and N 2(y) are closed linear complementary subspaces in H02(R(x)),
the sets A2(x) and B2(y) are closed linear complementary subspaces in L02(R(x)).
Further, we deduce from (13) that

c1 ≤ A(XT (x)) ≤ c2

and from (11) and (17) that

A(XT (x))B(YT (y)) = 1.

Hence, the optimization problems (21) and (22) satisfy the conditions of Lemma 2
and the result follows.

We proceed now with the proof of the main theorem of the paper.

PROOF OF THEOREM 1. For x > 0, denote by M∞(x) the family of semi-
martingales M such that for some ε = ε(M) > 0,

(87) X(x)(1 + εM) ∈ X(x) and X(x)(1 − εM) ∈ X(x),

where X(x) is defined in (2). If M ∈ M∞(x), then M is uniformly bounded and
has the initial value 0. Note that the set M∞(x) has the economic interpretation
as the set of uniformly bounded wealth processes with initial value 0 under the
numéraire X(x).

Similarly, for y > 0, denote by N ∞(y) the family of semimartingales N such
that for some ε = ε(N) > 0,

(88) Y (y)(1 + εN) ∈ Y(y) and Y (y)(1 − εN) ∈ Y(y),

where Y(y) is defined in (8). If N ∈ N ∞(y), then it is uniformly bounded and
N0 = 0.

The following lemma plays the crucial role in the proof.

1370 D. KRAMKOV AND M. SÎRBU

LEMMA 6. Assume the conditions of Theorem 1 and let x > 0 and y = u (x).
Then, the sets M∞(x) and N ∞(y) belong to H02(R(x)), and their respective clo-
sures in H02(R(x)) coincide with the sets M2(x) and N 2(y) defined in Section 1.

PROOF. From the definitions of the probability measure R(x) and the families
M∞(x) and N ∞(y), we deduce that if M ∈ M∞(x) and N ∈ N ∞(y), then, for

sufficiently small ε > 0, the processes 1 + εM and 1 − εM, 1 + εN and 1 − εN ,

(1 + εM)(1 + εN) and (1 + εM)(1 − εN) are supermartingales under R(x). This

clearly implies that the bounded processes M, N and MN are martingales un-

der R(x). Hence, M∞(x )thaantdMN∞∞(x(y))caoriencoirdtehsowgoitnhalthseubsespt aocfebsoiunnHde20d(Rst(oxc)h)a. stic
From (87), we deduce

integrals with respect to the process SX(x) defined in (19). Hence,

M∞(x) ⊂ M2(x).

Assumption 2 implies the existence of a strictly positive predictable one-dimen-
sional process h such that the stochastic integral

S = h dSX(x)

is well defined and is a locally bounded process. If L ∈ M2(x), then L ∈ H02(R(x))
and there is a predictable process H such that

(89) L = H dS.

As S is locally bounded, we deduce that L can be approximated in H20(R(x))
by bounded stochastic integrals with respect to S and, hence, by elements
from M∞(x). It follows that the closure of M∞(x) in H02(R(x)) coincides
with M2(x).

Taking into account orthogonality relations between M∞(x) and N ∞(y), we

deduce that

N ∞(y) ⊂ N 2(y).

To finish the proof, it remains to be shown that any L ∈ H02(R(x)) such that LN is
a martingale under R(x) for any N ∈ N ∞(y) is an element of M2(x) or, equiva-
lently, has the integral representation (89).

Denote by Q the family of equivalent local martingale measures for S which
have bounded densities with respect to R(x) and by Z the family of density
processes of these measures. We claim that

(90) Z − 1 ∈ N ∞(y), Z ∈ Z.

Indeed, if Z ∈ Z, then for any X ∈ X(1) we have that X Z is a local martingale
X(x)
under R(x) and, hence,

XZY (y) = X ZX(x)Y (y)
X(x)

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1371

is a local martingale under P. It follows that

(91) ZY (y) ∈ Y(y), Z ∈ Z.

Relation (90) is now implied by (91) and the fact that for any Z ∈ Z there is a
sufficiently small ε > 0 such that

1 ± ε(Z − 1) ∈ Z.

From (90) and the assumption that LN is a martingale under R(x) for all
N ∈ N ∞(y), we deduce that L is a martingale under all Q ∈ Q. The integral

representation (89) now follows from the well-known result by Jacka; see [5],

Theorem 1.1.

Given Lemma 6, the proof of the theorem is a rather straightforward conse-
quence of its “abstract version,” Theorem 2.

Let x > 0 and y = u (x). By C(x) and D(y), we denote the sets of positive
random variables which are dominated by the final values of the processes from
X(x) and Y(y), respectively. That is,

(92) C(x) = {g ∈ L0 : 0 ≤ g ≤ XT for some X ∈ X(x)},

(93) D(y) = {h ∈ L0 : 0 ≤ h ≤ YT for some Y ∈ Y(y)}.

By A∞(x) and B∞(y), we denote the sets of final values of elements from
M∞(x) and N ∞(y), respectively. That is,

(94) A∞(x) = {α ∈ L∞ : α = MT for some M ∈ M∞(x)},

(95) B∞(y) = {β ∈ L∞ : β = NT for some N ∈ N ∞(y)}.

With this notation, the value functions u and v defined in (5) and (7) take the form
(46) and (47).

According to Proposition 3.1 in [7], the sets C(1) and D(1) satisfy con-
ditions (43)–(45). It is easy to see that the sets C(x) and D(y) defined in
(92) and (93) are related to the sets A∞(x) and B∞(y) defined in (94) and (95)
in the same way as the corresponding sets in Section 2, that is, through formu-
las (51) and (52). Finally, from Lemma 6, we deduce that the respective closures
of A∞(x) and B∞(y) in L02(R(x)) are given by the sets A2(x) and B2(y) de-
fined in (85) and (86). In particular, these closures are complementary subspaces
in L20(R(x)). Hence, all the assumptions of Theorem 2 are satisfied.

From Theorem 2 we deduce all the assertions of Theorem 1, except the fact
that the products X(x)Y (y), X (x)Y (y) and X (x)Y (y) are martingales un-
der P. However, this result is an immediate consequence of the definitions of the
processes X (x) and Y (y).

1372 D. KRAMKOV AND M. SÎRBU

4. Counterexamples. This section is devoted to (counter)examples related to
our main results. In the first three examples we show that the assertions of Theo-
rem 1 might not hold true if one of the Assumptions 1 or 2 is not satisfied.

EXAMPLE 1. We show that the lower bound

xU (x) x > 0,
c1 ≤ − U (x) ,

in Assumption 1 cannot be removed without affecting the existence of the second-
order derivative v .

We start by choosing a continuous function φ : (0, ∞) → (0, ∞) such that

(96) φ(k) = 2k, k = 1, 2, . . . ,

∞

(97) φ(s) ds = ∞,

0

∞∞

(98) φ(s) ds dt < ∞

0t

and for some c2 > 0,

(99) tφ(t) > 1∞ φ(s) ds, t > 0.

c2 t

To construct such a function φ, we can start, for example, with the function

ψ (t ) = t 1 e−t , t > 0,

3/2

which satisfies (97)–(99), and then modify its values near integers so that (97)–(99)

still hold true and, in addition, (96) is satisfied.

We now define

y∞

V (y) = − φ(s) ds dt,

0t

U (x) = inf {V (y) + xy}.

y>0

Conditions (97) and (98) imply that U satisfies the usual assumptions of a utility
function, including Inada conditions (4), and that U and V are bounded:

−∞ < V (∞) = U (0) < V (0) = U (∞) = 0.

In addition, as

∞ V (y) = φ(y), y > 0,

V (y) = − φ(t) dt,

y

condition (99) is equivalent to the upper bound in Assumption 1:

− xU (x) < c2, x > 0.
U (x)

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1373

We now choose a probability space ( , F , (Ft )0≤t≤T , P), where the filtration

(Ft )0≤t≤T is generated by a Brownian motion W . On this probability space, we
1
consider a discrete random variable ξ which takes the values 2 , 1, 2, 3, . . . , n, . . .

and such that

(100) E[ξ ] = 1 and P[ξ = k] = 2−k for large k.

It is easy to see that we can construct a continuous stock market (thus trivially
satisfying Assumption 2) which is complete and such that the unique martingale
measure has the density ξ .

For this financial model, the dual value function has the representation

v(y) = E[V (yξ )], y > 0.

As V is bounded, so is v. It follows that v is continuously differentiable (see [8]
for a general version of this result) and

(101) v (y) = E[ξ V (yξ )], y > 0.

Using the Taylor expansion for the function V around ξ(ω) and (101), we deduce

(102) v(1 + ε) = v(1) + εv (1) + 1 ε2E[V (ηε)ξ 2],
2

where ηε(ω) ∈ [ξ(ω), (1 + ε)ξ(ω)]. From (96) and (100), we obtain

(103) E[V (ξ )ξ 2] = E[φ(ξ )ξ 2] = ∞.

Taking into account that V (ηε) → V (ξ ) a.s. we can use (103) and Fatou’s lemma
to get

lim v(1 + ε) − v(1) − εv (1) = ∞.
(1/2)ε2
ε→0

Hence v is not two-times differentiable at y = 1.

EXAMPLE 2. We show that the upper bound

xU (x) x > 0,
− U (x) ≤ c2,

in Assumption 1 is essential for the second-order differentiability of u.
In a similar way to Example 1, we can construct a bounded two-times continu-

ously differentiable utility function U such that the lower bound in Assumption 1
holds true, and

(104) −U (k) = 2k for large k.

We consider a one-period financial model with the stock price process S =

(S0, S1), where S0 = 1 and S1 is a discrete random variable that takes the val-
1
ues 2 , 1, 2, 3, . . . , k, . . . and satisfies the following conditions:

(105) P[S1 = k] = 2−k for large k,

(106) E[U (S1)(S1 − 1)] = 0.

1374 D. KRAMKOV AND M. SÎRBU

Note that, for this model, the set of nonnegative wealth processes with initial value
x is given by

(107) X(x) = {x + a(S − 1) : a ∈ [0, 2x]}.

From (106), we deduce that the optimal investment strategy for x = 1 is to buy
and hold one unit of the stock, that is,

(108) X(1) = S.

Indeed, if X ∈ X(1), then, by (106) and (107), we have

E[U (S1)X1] = E U (S1) 1 + a(S1 − 1) = E[U (S1)] = E[U (S1)S1].

Using the notation V for the convex conjugate to U , we deduce

E[U (X1)] ≤ E[V (U (S1)) + U (S1)X1]
= E[V (U (S1)) + U (S1)S1] = E[U (S1)],

thus proving (108). We point out that Assumption 2 is satisfied for this model at
x = 1 because

SX(1) = 1S = 1
, ,1
X(1) X(1) S

is bounded.
Assumptions (104) and (105) yield

(109) E U (S1) 1 + a(S1 − 1) 2 = −∞, a ∈ (−∞, +∞).

For any |ε| < 1 we know from (107) that

u(1 + ε) = E U 1 + ε + aε(S1 − 1)

=E U S1 + ε 1 + aε − 1 (S1 − 1) ,
ε

for some aε ∈ [0, 2(1 + ε)]. Using the Taylor expansion we obtain

u(1 + ε) = u(1) + εu (1) + 1 ε2E aε − 1 2
2 ε
U (ηε ) 1 + (S1 − 1) ,

where ηε is a random variable that converges to S1 as ε → 0. A subsequence
aε −1
argument for ε , together with (109) and Fatou’s lemma, now implies that

lim u(1 + ε) − u(1) − εu (1) = −∞.

ε→0 (1/2)ε2

Hence, u is not two-times differentiable at x = 1.

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1375

EXAMPLE 3. We show here that if Assumption 2 is violated, then u and v

might not be two-times differentiable.

We choose a one-period financial model with one stock, where S0 = 1 and S1
1 1
takes the values 2, 1, 2 , 4 , . . . and satisfies

(110) E1 =1 and E 1 < ∞.
S1 S12

We point out that, for this model, the set of nonnegative wealth processes with
initial wealth x is given by

(111) X(x) = {x + a(S − 1) : a ∈ [−x, x]}.

We choose two-times continuously differentiable utility function U which is
bounded above, which satisfies Assumption 1 and such that

(112) U (x) = 1 x = 1 , n ≥ −1,
, 2n
x

and

(113) E[U (S1)(1 − S12)] = 0.

Note that, by Assumption 1 and (112), the random variable |U (S1)S12| is bounded
and, therefore, the second inequality in (110) implies that

E[|U (S1)|(1 + S12)] < ∞.

It follows that

1 + S1 2 (S1) 1 + a(S1 − 1) 2
2
(114) EU (S1) = sup E U <∞

a

and, by (113), that the upper bound in (114) is attained at a = 1/2.
As in Example 2, we deduce from (110) and (112) that the optimal investment

strategy for x = 1 is to buy and hold one unit of the stock, that is,

X(1) = S.

We point out that

u (1) = E[X1(1)U (X1(1))] = E 1 = 1,
S1 S1

and that the process

SX(1) = 1S = 1
, ,1
X(1) X(1) S

is not sigma-bounded.

1376 D. KRAMKOV AND M. SÎRBU

For ε > 0, we have S + ε 1+S ∈ X(1 + ε), so
2

(115) u(1 + ε) ≥ f (ε) := E U S1 + ε 1 + S1 .
2

From [9], Lemma 1, we deduce that the function f has first and second derivatives
from the right at 0 given by

(116) f+(0) = E U (S1) 1 + S1 = 1 = u (1),
2

(117) f+(0) = E U (S1) 1 + S1 2
2
.

Taking into account (114), (115), (116) and (117), we conclude that

(118) lim inf u(1 + ε) − u(1) − εu (1) ≥ sup E U (S1) 1 + a(S1 − 1) 2 .

ε0 (1/2)ε2 a

From (111), we conclude that, for ε > 0,

X1(1 − ε) − X1(1) = 1 + aε (S1 − 1) where aε ≥ 1.
−ε

Now, according to the Taylor expansion,

u(1 − ε) − u(1) + εu (1) =E U (ξε) 1 + aε(S1 − 1) 2 ,

(1/2)ε2

for ξε ∈ [X1(1), X1(1 − ε)]. Since U is assumed to be continuous, using a subse-
quence argument for aε and Fatou’s lemma, we obtain

(119) lim sup u(1 − ε) − u(1) + εu (1) ≤ sup E U (S1) 1 + a(S1 − 1) 2 .

ε0 (1/2)ε2 a≥1

From (118) and (119), taking into account that the supremum in (114) is strictly
1
attained at a = 2 , we conclude that u does not have a second derivative at x = 1.

EXAMPLE 4. We construct a very simple financial model that satisfies the

conditions of Theorem 1, and such that the derivative processes X (x) and Y (y)

defined in Section 1 are negative and equal to 0 with a positive probability.

We choose a one-period model such that S0 = 1, and S1 takes the val-
1 1 1
ues 8 , 4 , 2 , 2 with positive probabilities such that

(120) 1
E = 1.

S1

We choose a bounded two-times continuously differentiable utility function U that
satisfies Assumption 1 and such that

(121) U (x) = 1 x = 1 1 1
, , , , 2,
x 842

(122) E U (S1) 1 + 4 (S1 − 1) (S1 − 1) = 0.
3

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1377

Using (120) and (121), we obtain that the optimal investment strategy for the initial
capital x = 1 is to buy and hold one unit of the stock, that is,

X(1) = S.

From (122), we deduce that the function

f (a) = E U (S1) 1 + a(S1 − 1) 2 , a ∈ (−∞, +∞),

attains its maximum at a = 4/3. It follows from Theorem 1 that the derivative
process X (1) is given by

X (1) = 1 + 4 (S − 1).
3

We now have only to see that

P[X1(1) = 0] = P S1 = 1 > 0,
4

P[X1(1) < 0] = P S1 = 1 > 0.
8

EXAMPLE 5. To motivate the current formulation of Assumption 2 in terms of

sigma-bounded processes we construct a complete financial model with a bounded

price process S, where, in the case of a logarithmic utility function, the semimartin-
gale SX(x) defined in (19) is not locally bounded.

Let N be a Poisson process with intensity 1 defined on a probability space
( , F , (Ft )0≤t≤T , P), where the filtration (Ft )0≤t≤T is generated by N . We
choose a continuous function φ : (0, ∞) → (0, ∞) such that

(123) lim φ(t) = ∞,

t0

(124) T

φ(t) dt = 1/2,

0

and define the process Z as follows:

t 0≤t ≤T.

Zt = 1 + φ(u) d(Nu − u),

0

From (124) we conclude that Z is a martingale and that

(125) Zt ≥ 1/2, 0 ≤ t ≤ T .

The process Z is clearly sigma-bounded. However, due to (123) it is not locally
bounded. In fact, it is easy to see that the only stopping time τ such that Z is
bounded on [0, τ ] is τ = 0. We define the price process for the stock as

S= 1.
Z

From (125) we deduce that S is a nonnegative price process bounded from above
by two. Standard arguments show that Z is the density process of the unique mar-
tingale measure and, hence, the model is complete.

1378 D. KRAMKOV AND M. SÎRBU

Consider now the problem of expected utility maximization with the logarithmic
utility function U (x) = log(x). As

U (x) = 1/x, x > 0,

we conclude that the optimal investment strategy for any initial capital is to invest
it into the stock, that is,

X(x) = xS.

The price process of the traded securities (the bond and the stock) under the
numéraire X(x) becomes

SX(x) = 1S = 1
, (Z, 1)
xS xS x

and, as we argued above, is sigma-bounded but is not locally bounded.

APPENDIX A: ON SIGMA-BOUNDED SEMIMARTINGALES

In this section we explore the concept of sigma-bounded semimartingales intro-
duced in Definition 1 and present convenient sufficient conditions for the validity
of Assumption 2. We start with two simple observations. The first one, for which
we skip the proof, is a characterization of locally bounded semimartingales. The
second one contains a similar description of sigma-bounded semimartingales. As
before, we work on the standard filtered probability space ( , F , (Ft )0≤t≤T , P).

LEMMA 7. For a d-dimensional semimartingale R, the following conditions
are equivalent:

1. R is locally bounded.
2. R has the integral representation

(126) t 0≤t ≤T,

Rt = R0 + Hu dSu,

0

where S is a uniformly bounded semimartingale and H is an increasing, pre-
dictable and S-integrable process.
3. R is dominated by some predictable increasing process K, that is,

(127) Rt ≤ Kt , 0 ≤ t ≤ T .

LEMMA 8. For a d-dimensional semimartingale R, the following conditions
are equivalent:

1. R is sigma-bounded.
2. R has the integral representation (126), where S is a uniformly bounded semi-

martingale, and H is a predictable and S-integrable process.
3. R is dominated by some predictable process K, that is, (127) holds true.

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1379

PROOF. To simplify notation, we assume that all processes appearing in
Lemma 8 are one-dimensional. Of course, this assumption does not restrict any
generality. The implication 1 ⇒ 2 is straightforward. The implication 2 ⇒ 3 fol-
lows from

|Rt | ≤ |Rt−| + |Ht || S| ≤ |Rt−| + 2c|Ht | := Kt ,

where c is a constant dominating the bounded semimartingale S. Finally, if (127)

holds true for a predictable process K, then | R| ≤ 2K and, hence, the stochastic

integral dR is locally bounded. This proves 3 ⇒ 1.
1+K

We now state a convenient sufficient condition on the filtered probability space
( , F , (Ft )0≤t≤T , P) that implies the sigma-boundedness property for any semi-
martingale. Recall that a local martingale N is called purely discontinuous if NM

is a local martingale for any continuous local martingale M.

ASSUMPTION 4. There is a d-dimensional local martingale M such that any
bounded, purely discontinuous martingale N is a stochastic integral with respect
to M, that is,

(128) t 0≤t ≤T,

Nt = N0 + Hu dMu,

0

for some predictable and M-integrable H .

REMARK 3. As the proof of Theorem 3 shows, Assumption 4 is invariant with
respect to an equivalent choice of reference probability measure. More precisely,
if it holds true under P, then it is also satisfied under any probability measure Q
that is absolutely continuous with respect to P: Q P. Also, Theorem 3 implies
that the integral representation (128) holds true for any (not necessarily bounded)
purely discontinuous local martingale N .

THEOREM 3. Assume that Assumption 4 holds true. Then any semimartin-
gale X defined on the filtered probability space ( , F , (Ft )0≤t≤T , P) is sigma-
bounded.

The proof of Theorem 3 relies on Proposition 1 below, which is a result of
independent interest. For a d-dimensional semimartingale R we denote by I(R)
the set of stochastic integrals with respect to R, that is,

I(R) = X : X = H dR for some predictable H ,

and by I∞(R) the set of bounded elements of I(R):
I∞(R) = {X : X ∈ I(R) and is bounded}.

1380 D. KRAMKOV AND M. SÎRBU

We use the standard notation H1 = H1(P) for the space of uniformly integrable
martingales under P such that

M H1 := E[MT∗ ] < ∞,
where MT∗ = sup0≤t≤T Mt .

PROPOSITION 1. Let R be a d-dimensional semimartingale. There exists a
d-dimensional bounded semimartingale S ∈ I∞(R) such that

I∞(R) = I∞(S).

PROOF. By changing, if necessary, the probability measure P to an equivalent
one, we can assume without any loss of generality that R is a special semimartin-
gale, that is,

R = R0 + M + A,

where M is a local martingale and A is a predictable process of finite variation.
Making the observation that there exists ϕ > 0 such that M = ϕ dM ∈ H1 and

A = ϕ dA has integrable variation, we can also assume that

M ∈ H1 and E T

(129) dAt < ∞.

0

Let [M, M] = di=1[Mi, Mi]. We would like to point out that our definition of
[M, M] differs from the matrix-valued process often used in the literature. De-
noting by C the (one-dimensional) compensator of [M, M]1/2, we now define the

measure µ on the predictable σ -field of [0, T ] × by

dµ(t, ω) = dCt (ω) + dAt (ω) dP(ω).

From (129) we conclude that the measure µ is finite:

µ([0, T ] × ) < ∞.

Denote by A the set of predictable processes with values in the set of d × d
symmetric and positive semidefinite matrices such that A dR is well defined and
locally bounded. We claim that there is some A ∈ A such that

(130) rank(A) ≥ rank(A) ∀ A ∈ A,

where rank(A) denotes the rank of the matrix A (and the inequality holds µ-a.s.).
Let (An)n≥1 be a sequence in A such that

(131) lim rank(An) dµ = sup rank(A) dµ.

n→∞ [0,T ]× A∈A [0,T ]×

We claim that the elements of this sequence can be chosen such that

(132) An + An R ≤ 1

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1381

and

(133) rank(An+1) ≥ rank(An).

Indeed, the condition (132) is easy to fulfill, while in order to satisfy (133), it is
sufficient to pass from (An)n≥1 to the sequence (Bn)n≥1 defined by

B1 = A1,

Bn+1 = BnI{rank(Bn)≥rank(An+1)} + An+1I{rank(Bn)<rank(An+1)}.

Assuming (132) and (133), and denoting by N the predictable process

N = max rank(An)

n≥1

we now define A as

∞

A = AnI{rank(An)=N,rank(Ak)<N,k=1,2,...,n−1}.

n=1

We deduce that A is an element of A such that the upper bound on the right-hand
side of (131) is attained. As for any A ∈ A the sum A + A is in A, and

rank(A + A) ≥ rank(A),

it follows that A is also a maximal element in the sense of (130).
We can choose A so that (in addition to (130)) the stochastic integral

(134) S = AdR

is bounded (not only locally). Denote by H the set of d-dimensional predictable
processes H such that H dR is locally bounded. To complete the proof, it re-
mains to be shown that any H ∈ H admits the representation

(135) H = AG

for some predictable d-dimensional process G. Indeed, in this case,

H dR = H, dR = AG, dR = G, A dR = G dS.

To prove (135) we use the fact that any predictable d-dimensional process H
can be decomposed as

(136) H = AG + F,

where G and F are predictable d-dimensional processes such that F ∈ ker(A)
[ker(A) denotes the kernel of the matrix A]. We have to show that for H ∈ H the
process F in (136) equals zero.

1382 D. KRAMKOV AND M. SÎRBU

Multiplying, if necessary, both sides of (136) by a strictly positive predictable
process, we can assume that F ∈ H and F ≤ 1. In this case, the matrix B defined
by

Bij = F iF j , 1 ≤ i, j ≤ d,

belongs to A. Hence, A + B ∈ A. However, as F ∈ ker(A)

rank(A + B) = rank(A) + I{F =0},

and the equality of F to zero follows from the maximality property (130) for A.

PROOF OF THEOREM 3. By Proposition 1, we can assume, without loss of
generality, that the local martingale M appearing in Assumption 4 is bounded. We
can also assume that M is purely discontinuous. We shall maintain these assump-
tions about M throughout the proof.

Let Z be a bounded martingale orthogonal to M, that is, ZN is a martingale and

Z = Zc + Zd, Z0d = 0,

is its decomposition into the sum of the continuous martingale Zc and the purely
discontinuous martingale Zd . As Zd is locally bounded, it is a stochastic integral
with respect to M. Since Zd is orthogonal to M, we deduce that Zd = 0.

We have thus proved that any bounded martingale orthogonal to M is contin-

uous. It follows that any purely discontinuous local martingale is a stochastic in-

tegral with respect to M; see, for example, Theorem 1.1 in [5]. In particular, any

purely discontinuous local martingale and, therefore, any special semimartingale,

is sigma-bounded.
Since a semimartingale becomes a special semimartingale under an equivalent

probability measure, it remains to be shown that Assumption 4 holds true under
any P P. Let

dP 0≤t ≤T,
Zt = E dP Ft ,

be the density process of P with respect to P and L = dZ/Z− be the stochastic
logarithm of Z. It is well known that any locally bounded local martingale N under
P has the representation

(137) N = N − N,L ,

where N is a locally bounded local martingale under P, and N, L is the pre-

dictable covariation between P-local martingales N and L. If, in addition, N is
purely discontinuous, then N is also purely discontinuous. Assumption 4 implies
the existence of integral representation (128) for N in terms of M. It follows that N

TWO-TIMES DIFFERENTIABILITY IN OPTIMAL INVESTMENT 1383

has the integral representation (with the same integrand H ) in terms of P-local
martingale M defined by

M = M − M,L .

Hence, Assumption 4 is satisfied under P.

The next theorem characterizes local martingales that are sigma-bounded.

THEOREM 4. For a local martingale R, the following assertions are equiva-
lent:

1. R is sigma-bounded. dQ
dP
2. If Q is a probability measure such that ∈ L∞ and any X ∈ I∞(R) is a

martingale under Q, then any X ∈ I(R) ∩ H1 is a martingale under Q.

3. The closure of I∞(R) in H1 coincides with I(R) ∩ H1.

PROOF. The implications 1 ⇒ 3 and 3 ⇒ 2 are straightforward. To prove the
remaining implication 2 ⇒ 1 we assume that R ∈ H1. Of course, this does not

restrict any generality.

Let S be the bounded semimartingale (in fact a martingale) given by Proposi-

tion 1. Consider a probability measure Q such that S is a local martingale under Q
dQ
and dP ∈ L∞. We have that any element in I∞(R) = I∞(S) is a martingale un-

der Q. According to condition 2, this implies that R is a martingale under Q. From

Jacka’s theorem (see [5], Theorem 1.1) we deduce that R is a stochastic integral

with respect to S. The semimartingale S being bounded, we have thus proved that

R is sigma-bounded.

We conclude this section with an easy corollary of Theorem 3 showing, in par-

ticular, that complete financial models satisfy Assumption 2. Hereafter, we con-

sider the financial model with d-dimensional price process S and the nonempty

family of equivalent martingale measures Q introduced in Section 1. Recall that
the model is called complete if any f ∈ L∞ can be represented as the terminal
value of a bounded wealth process, that is, f = XT for some process X such that
X − X0 ∈ I∞(S). It is well known (see [3]) that the model is complete iff Q is a
singleton.

PROPOSITION 2. Assume that the financial model is complete. Then any semi-
martingale X defined on the filtered probability space ( , F , (Ft )0≤t≤T , P) is
sigma-bounded.

PROOF. Denote by Q the unique element of Q. Assume first that S is a local
martingale under Q. As the model is complete, any bounded martingale under Q

1384 D. KRAMKOV AND M. SÎRBU

belongs to I∞(S). Hence Assumption 4 holds true and the result follows from
Theorem 3.

In the general case, we use Proposition 1 to find a bounded semimartingale
S ∈ I∞(S) such that

I∞(S) = I∞(S).

We have that S is a bounded martingale under Q and that any bounded martingale
under Q belongs to I∞(S). The result again follows from Theorem 3.

Acknowledgments. We want to thank an Associate Editor and the referees for
their comments and remarks. Part of the work was done while Mihai Sîrbu was a
graduate student at Carnegie Mellon University.

REFERENCES

[1] DELBAEN, F. and SCHACHERMAYER, W. (1998). The fundamental theorem of asset pricing
for unbounded stochastic processes. Math. Ann. 312 215–250. MR1671792

[2] GOURIEROUX, C., LAURENT, J. P. and PHAM, H. (1998). Mean-variance hedging and
numéraire. Math. Finance 8 179–200. MR1635796

[3] HARRISON, J. M. and PLISKA, S. R. (1983). A stochastic calculus model of of continuous
trading: Complete markets. Stochastic Process. Appl. 15 313–316. MR0711188

[4] HIRIART-URRUTY, J. and LEMARÉCHAL, C. (1993). Convex Analysis and Minimization Al-
gorithms. Springer, Berlin. MR1261420

[5] JACKA, S. D. (1992). A martingale representation result and an application to incomplete fi-
nancial markets. Math. Finance 2 239–250.

[6] KALLSEN, J. (2004). σ -localization and σ -martingales. Theory Probab. Appl. 48 152–163.
MR2013413

[7] KRAMKOV, D. and SCHACHERMAYER, W. (1999). The asymptotic elasticity of utility func-
tions and optimal investment in incomplete markets. Ann. Appl. Probab. 9 904–950.
MR1722287

[8] KRAMKOV, D. and SCHACHERMAYER, W. (2003). Necessary and sufficient conditions in the
problem of optimal investment in incomplete markets. Ann. Appl. Probab. 13 1504–1516.
MR2023886

[9] KRAMKOV, D. and SÎRBU, M. (2006). The sensitivity analysis of utility based prices and the
risk-tolerance wealth processes. Ann. Appl. Probab. To appear.

[10] KRYLOV, N. V. (1980). Controlled Diffusion Processes. Springer, New York. MR0601776

DEPARTMENT OF MATHEMATICAL SCIENCES DEPARTMENT OF MATHEMATICS
CARNEGIE MELLON UNIVERSITY COLUMBIA UNIVERSITY
5000 FORBES AVENUE 2990 BROADWAY
PITTSBURGH, PENNSYLVANIA 15213 NEW YORK, NEW YORK 10027
USA USA
E-MAIL: [email protected] E-MAIL: [email protected]